A Human-Inspired Controller for Fluid Human-Robot Handovers

Jose R. Medina∗ Felix Duvallet∗ Murali Karnam Aude Billard

Abstract—Handovers are seamless events that occur fre-quently and naturally between people. Although previous workshave focused on the design of robot handover controllerssynthesizing one of the phases of a handover (approaching,passing, or retracting) in an isolated manner, we take a differentapproach and treat the handover as a single continuous entity.In this paper we present a novel bi-directional and human-inspired handover controller using insights we learned fromhuman-human experiments. We model the dynamics of thehandover in a continuous and time-independent way yieldingrobust and fluid behavior. We implemented our approach on arobot platform consisting of a 7-DoF robotic arm with a 16-DoF humanoid hand. Our results show that resulting human-robot handovers are smooth and reduce internal forces with thehuman compared to traditional switching approaches.

I. INTRODUCTION

A handover occurs anytime two parties exchange an objectwithout the use of an external tool (such as a table). Thesefast and seamless events occur naturally between people, ina wide range of environmental conditions with and withoutconstraints, and for objects ranging in size and mass. Tobecome useful partners in human-robot teams, robots willneed to approach this level of capability, and be able toperform fast natural handovers of arbitrary objects as bothgiver and receiver.

Once a handover begins, it generally consist of three parts:an approach phase where one or both parties move to thehandover location (where the object is reachable by both),a passing phase where the object is physically transferredfrom the giver to the receiver, and a retraction phase wherethe giver and receiver move away from each other. Whilemuch work exists that attempts to understand the dynamicsof this handover process in humans, to date robot controllersattempting to replicate these processes have fallen short ofproviding a complete handover implementation that is fluidand seamless. One explanation for this gap is that priorstudies that replicated these different phases generally didso by considering each phase in isolation (e.g., movingthe arm in a predictable way, or controlling the grip forceduring object transfer but without any robot arm motion).This results in a handover that is un-naturally sequential: thehandover proceeds from one phase to the next, rigidly waitingfor a specific condition before transitioning to the next phase(e.g., a waiting for a force trigger before completely openingthe hand).

We take a different approach, treating the handover asa single entity where we design jointly the motions of the

∗The first two authors contributed equally to this paper. All authors arewith the LASA laboratory of the Ecole Polytechnique Federale de Lausanne,Switzerland. {firstname.lastname}@epfl.ch

Fig. 1. Our goal is to enable natural and fluid handovers of objects betweenpeople and robots. Using studies of human-human handovers, we analyzedthe role of motion, grip forces, and interaction wrenches to model thedynamics of fluid handovers. We then developed a controller that estimatesthe location of the handover, moves the arm towards the handover pose ina natural way, and releases the object in an intuitive manner. All of theseoccur in a continuous and phase-less way.

complete robot (in our case, arm and hand) to perform a fluidhandover. Our controller is human-inspired, using insightsacquired from human-human handover experiments. Theresulting handover is natural, bi-directional, and adaptableto many object shapes and weights.

The components for our approach, shown in Figure 1,are discussed in the rest of this paper. After reviewing priorwork in the area of handovers (Section II), we present themathematical formulation of the problem in Section III. Wethen describe our human-human handover experiments andthe insights derived from them in Section IV; importantly weshow that people control both their hand and arm during allhandover phases depending on the proportion of the objectload they support, implying that a robot controller shouldunderstand, decompose, and robustly react to interactionforces in order to perform a fluid handover. We describe sucha human-inspired controller in Section V, and present resultsfrom human-robot handover experiments in Section VI.

II. RELATED WORK

Given their frequency and importance in day-to-day life,it is not surprising that object handovers are well studied byresearchers. Insights about handovers come from two majorareas of literature: observations of human-human handoversthat attempt to understand the complex dynamics underpin-

ning handovers, and user studies (with a robot implemen-tation) that attempt to replicate (and analyze) some of thecomplex dynamics on a robot.

Across these broad areas of study, the major types ofstudies by researchers are studies into the forces exertedduring handovers (e.g., the grip forces exerted on the objectby both parties) and the motions of the handover participants(e.g., trajectories to the handover location).

A. Studies of handover forcesStudies on forces exerted during handover analyze the

relationship between grip force (the amount of force appliedon the object by each participants’ fingers) and load force(the proportion of the object’s weight sustained by eachparticipant). For example, Mason and MacKenzie [1] foundthat human-human handovers involve a gradual modulationof the grip force while both participants implicitly share (andtransfer) the load force. The main coordination mechanismis haptic feedback in order to coordinate the transfer forcerates; the grip force increases for the receiver and decreasesfor the passer.

Another consistent finding in prior studies is that the timerequired to transfer the object from the passer to the receiveris between 300 and 500 ms [1, 2, 3]. Even Endo et al.[3], who outfitted the receiver with a glove in some of thetrials to attenuate the tactile information, found that while theglove did increase the duration of the average contact periodslightly (from 324 ms to 334 ms), it did not have a significanteffect on the grip force profiles.

Two additional useful findings by Chan et al. [2] are thatthe passer has a “post-unloading” phase towards the endof the handover, where the giver applies a positive gripforce even though their load force is approximately zero.This, according to the authors, implies that the giver takesresponsibility for the safety of the object. Additionally, theirstudies found that the receiver adjusts the load transfer ratedepending on the weight of the object, and concluded thatthe receiver is in charge of the handover timing. The recom-mendations resulting from this study were used to develop alinear controller that varied the robot’s grip force according tothe sensed load force during a robot-to-human handover [4].However, the robot controller implementations based on thesestudies focus solely on the grip force modulation, with no armmotion during the handover.

These studies have important implications for designingrobot handover controllers:• the load force drives the dynamics of the passing phase

of handovers,• handovers must be fast and respond quickly to changes

in the load force.This means our robot must be able to quickly measure theproportion of the object’s load it supports, and quickly adaptits grip force on the object during the passing phase of ahandover.

B. Studies of handover motionsStudies on handover motions for robots fall into two

broad categories: planning-based approaches that optimize

a complete trajectory, and controller-based approach thatcontinuously optimize some controller input.

Planning-based approaches formulate one or more partsof the handover as a search problem. For example, suchapproaches have been applied to finding configuration of therobot arm the agree with human preferences [5], plans involv-ing handovers in constrained environments [6], or handoverlocations that optimize the shared effort of the human androbot [7].

Controller-based approaches for robot handovers also exist,and generally attempt to reach a desired goal position [8],often by following a desired trajectory [9].

While the implementations in these studies resulted innatural trajectories to the handover location, the robot stopsmoving before transitioning to the passing phase (once bothparties make contact with the object) and in general theseapproaches did not address the handover release strategy. Ourhuman studies show that people move slightly during the in-contact (passing) phase, and thus it is important for robotsto consider the entire handover as one fluid motion, withouthard transitions between the handover phases.

III. HANDOVER FORMULATION

We consider the problem of handing over an object fromone party to another. We assume that each party consists ofan arm and a hand, that the object is rigid and its physicalproperties are known, and that initially one of the parties isgrasping the object. In our setting, we do not consider thesignaling problem and assume that the intention to performa handover has already been established. We consider ahandover successful when the object load is transferred fromthe giver to the receiver without being supported anywhereelse (e.g., falling on the ground or placed on the table). Inthe following, subscripts g, r will refer to the giver and thereceiver respectively.

We are interested in modeling the motion of the giverand receiver as they move towards the handover location (aphase we will informally refer to as the approach phase) andtransfer of the object’s load (similarly, the passing phase). Wedenote the arm end effector poses of the giver and receiverby xg =

[pTg o

Tg

]T,xr =

[pTr o

Tr

]T, where pg,pr ∈ R3

and og,or ∈ R4 are the Euclidean positions and orientationsin the axis-angle representation respectively. We also denotethe hand configurations by xh,g,xh,r ∈ Rh.

We represent the rigid body dynamics of an object o withmass mo ∈ R by

Mo(xo)xo + Coxo + go = uo , (1)

where xo =[pTo ω

To

]Twith po,ωo ∈ R3 are respectively

the translational and rotational velocity of its center ofmass (expressed in the inertial frame), the mass matrixis Mo(xo) = diag{moI3 , Jo(xo)} where Jo(xo) ∈ R3×3

is the inertia tensor, uo =[fTo τ

To

]T ∈ R6 is the resultingwrench applied on the object with fo, τ o ∈ R3 denotingthe force and torque respectively, go =

[mog

T 0T3]T

is the

wrench on the object due to gravity, and

Co =

[03×3 03×303×3 ωo × Jo(xo)

]represents Coriolis and centrifugal terms.

When one of the parties is grasping the object, the appliedwrench on the object uo corresponds to the sensed wrenchat the wrist us considering the kinematic constraints of thegrasp, i.e.

uo = Gus ,

where G usually denotes the grasp matrix. It is given by

G =

[I3 03×3P I3

],

where P =[p]× is the cross product matrix for p, the vector

from the object’s center of mass to the end-effector position.When two parties g and r are grasping the object, the

object-centered dynamics become

uo = Ggus,g +Grus,r .

Note that we are considering the grasp of the giver and thereceiver as a single wrench input and we are not modeling thewrenches due to each finger/contact point, as typically donein grasping. In this case, we can decompose the wrench inputof each party (denoted by i) as

us,i = G−1i (Amot,iumot,o +Aload,igo) + uint,i , (2)

where uint,i are the internal wrench components be-tween the giver and the receiver inducing no object mo-tion nor compensating gravity wrenches, and the matri-ces Amot,i, Aload,i ∈ R6×6 allocate wrench components induc-ing object motion umot,o =Mo(xo)xo + Coxo and due togravity go respectively and fulfill

umot,o = (Amot,g +Amot,r)umot,o

go = (Aload,g +Aload,r)go . (3)

They are given by

Amot,i =

[βf,iI3 03×303×3 βτ,iI3

]Aload,i =

[αiI3 03×303×3 03×3

], (4)

where αi, βf,i, βτ,i ∈ [0, 1] and∑i

αi = 1,∑i

βf,i = 1,∑i

βτ,i = 1. This unusual wrench distribution separating

wrenches due to motion and gravity is necessary for han-dovers, as they are characterized by the transfer of theobject’s load.

Since our decomposition directly represents the proportionof the load share supported by the giver and receiver inEquation (3), we are able to represent the overall objectiveof transferring the load from the giver to the receiver as anevolution from αg = 1.0 (the giver supports the object’s load)to αg = 0.0 (the receiver now supports the load). Therefore,

Fig. 2. Experimental setup used in the human-human experimentsconsisting of a CyberGlove with TekScan patches, OptiTrack markers, andan instrumented object with a force/torque sensor.

the problem considered in this work consists of the designof function

xgxrxh,gxh,r

= f

xgxrxh,gxh,rαg

(5)

that describes the motion of both parties fulfilling objec-tive lim

t→∞αg(t)→ 0.

IV. INSIGHTS FROM HUMAN-HUMAN HANDOVERSTUDIES

To better understand the dynamics at play during a fluidhandover, we collected data from human-human handoversof a sensorized object between two participants such as theone shown in Figure 2. We then analyzed these trials toextract insights about the motion of the giver and receiverduring the entire handover process, as well as the interactionforces exerted on the object while it is being transferred.This study serves two key purposes. First, it enables us toinfer the important variables and structural constraints thatunderlie the handover dynamics of Equation (5). Second, theyprovide a corpus of training data with which we will learnthe parameters for the chosen model.

In these studies, two participants exchanged a sensorizedobject consisting of two handles joined together with ATXNano 25 force/torque sensor. Each handle has a massof mhandle = 0.28 [kg] and the object has a total massof mo = 2mhandle [kg]. We also placed an OptiTrack markeron the object, and recorded its pose during the entire han-dover. One participant was instrumented with a data glove(the CyberGlove with TekScan patches on the fingers andpalm) and an OptiTrack marker attached to their forearm.The second participant had an OptiTrack marker attached totheir wrist but was not instrumented otherwise.

Both participants took turns acting as the giver and re-ceiver. For each handover, one of the participants grasped

x

y

z

x

y

z

Pose receiver at handover

Pose giverat handover

x

y

z

Inertial frame

Fig. 3. We represent all poses in the frame of the configuration at thehandover time. This enables us to represent handovers in a generalizableframe of reference.

the object with a power (whole-hand) grasp, waited for the“go” signal, and then handed the object over to the otherparticipant. The only grasping constraint was that both partiesshould grasp the object on opposite sides of the force/torquesensor (we did not constrain users to use precision grasps).When the handover was complete we stopped recording datafor that trial and moved to the next handover. We collected atotal of 25 trials of handovers in both directions, with threedifferent participants holding the object both horizontally andvertically.

A. Data Processing

We processed the pose data to smooth it and compute itsderivatives using a Savitzky-Golay filter, and determined thehandover poses for both participants (the pose at the timewhen the participants are closest to each other), xg,ho for thegiver and xr,ho for the receiver, as shown in Figure 3. Thesehandover poses provide a way to represent the handoverin a generalizable frame of reference with respect to theobject pose at the handover, and for notational simplicityin the remainder of this analysis the poses of both thegiver xg and the receiver xr will be expressed relative totheir respective handover pose, such that pi,ho = 0 andtherefore ‖pi − pi,ho‖ = ‖pi‖ represents the distance to thehandover pose.

The force/torque sensor enables us to estimate both theinternal wrench applied on the object uint, as well as theload share αi. The wrench captured by the force/torquesensor us,o expressed in the inertial frame is

us,o = us,r + uhandle ,

or considering the other side of the sensor

us,o = −(us,g + uhandle) ,

where uhandle is given by (1) but with mass mhandleand us,i = Gius,i. Given observed object poses xo, graspmatrices Gi and the object’s physical properties, the com-putation of uhandle and us,i is straightforward. To estimatethe load share, the computation of the wrench distributionmatrices from (2) is necessary. The parameters βf,i and βτ,iconcerning the motion-inducing wrenches are given by theleast squares solution of ‖us,i −Amot,iumot,i‖ fulfilling con-

straints from (4), yielding

βf,i = max

{min

{fT

s,ifmot,o

‖fmot,o‖2, 1

}, 0

},

βτ,i = max

{min

{τTs,iτmot,o

‖τmot,o‖2, 1

}, 0

}. (6)

This solution follows from applying Gauss’s principle of leastconstraint for multibody systems [10]. The load share iscomputed similarly although subtracting a priori the motion-inducing wrench components already considered in the com-putation of βf,i, i.e.

αi = max

{min

{(fs,i − βf,ifmot,i)

Tg

‖g‖2, 1

}, 0

}. (7)

Given the results from Equations (6) and (7), the internalwrench uint,i follows straightforwardly from Equation (2).Note that for computing αi, the only information necessaryabout the object’s physical properties is its mass.

To estimate the grip force applied by the instrumentedparty, we use the Tekscan patches located on the hand tocompute the total pressure applied by the participant on theobject. For simplicity, in our analysis the estimated gripforce ui is given by the sum of the measurements acrossall Tekscan patches.

The CyberGlove measures the 22-dimensional joint config-uration of the hand of the instrumented participant (similarlyto the grip force, we can compute the hand configurationfor only one participant). For simplicity, in our analysis weuse this information to compute the derivative of the sumacross all measurements of the hand configuration, yieldingapproximately xh,i.

B. Data Analysis and Important Insights

After processing the data, we analyzed it for useful insightswhich will determine the important signals that will be usedby our controller. In order to identify potential dependenciesbetween all variables involved in the handover we firstperformed a Granger causality test [11]. This is a statisticalhypothesis test that evaluates if the addition of a new timeseries component improves the prediction for another timeseries, compared to using the past values alone. In oursetting, if some variable Granger-causes (G-causes) one of thetarget variables involved in the handover (such as the desiredtrajectory or the grip force), then it might be beneficial toinclude it as an input in the controller. We used the MVGCMatlab Toolbox [12].

1) Insight: The giver and receiver are coupled: We eval-uated G-causality during the motion phases (approach andretract), between the distance of the giver’s and receiver’send-effectors to the handover position (‖pg‖ and ‖pr‖ re-spectively), and the change in the hand configuration xh.Our results, shown in Figure 4, indicate a dependency fromthe giver’s motion to that of the receiver’s. This suggests thatthe giver leads the motion during the approach phase, and thereceiver’s motion is coupled to it.

‖pg‖ ‖pr‖ ‖pg‖ ‖pr‖

‖pg‖ ‖pg‖

‖pr‖ ‖pr‖

from from

to toFig. 4. Granger causality (left: score, right: statistical significance) for thedistance of both end-effectors to the handover position during the approachphase. A darker square indicates higher causality.

ug αg uint,xuint,yuint,zxh,g ug αg uint,xuint,yuint,zxh,g

ug ug

αg αg

uint,x uint,x

uint,y uint,y

uint,z uint,z

xh,g xh,g

from from

to to

Fig. 5. Granger causality between the grip force of the giver ug , loadshare αg , the 3-dimensional internal wrench uint, and the hand configurationof the giver xh,g during the handover. A darker square indicates highercausality score. The red rectangle highlights the scores of all variables G-causing the grip force ug .

2) Insight: The load share drives the evolution of thepassing phase: Since the grip force of the giver ug willbe the key control input regulating the load transfer duringthe handover, we are interested in studying its relationshipto the load share αg , the internal wrenches sensed by thegiver uint (we hereafter drop the g index for simplicity) whichare not part of the load force, and the hand motion of thegiver xh,g . The G-causality results for the passing phase,shown in Figure 5, indicate that the two highest scores arefor the internal wrench in the y direction uint,y and the loadshare αg . Since in our experiments uint,y is aligned withthe direction of motion, it thus corresponds to pushing andpulling on the object by the participants in the direction ofthe handover.

To investigate this relationship further, we analyzed gripforce as a function of the load share. We find that thegrip force increases almost linearly as a function of theload share, as shown in Figure 6. This result suggests thathumans modulate grip force as a function of the amount ofperceived load share, and aligns with prior studies modelingthe relationship between the two variables [1, 4].

3) Insight: Both parties move during the passing phase:In addition, it is also interesting to study arm motion duringthe passing phase and its potential role in the load transfer.More precisely, motions in the z-axis are especially relevantas they directly influence the load force. Figure 7 showsthe difference between the giver and the receiver positions

0 0.2 0.4 0.6 0.8 10

200

400

600

800

Load share αg

Gri

pfo

rceug

Fig. 6. Relationship between the giver’s grip force ug as a function of theload share αg during handovers. We show both the raw points (plus) anda fit using Gaussian Process Regression, indicating the grip force decreasesalmost linearly as the load share decreases.

0 0.2 0.4 0.6 0.8 1

−4

−2

0

2

4

6

·10−2

Load share αg

pg−pr

(z-a

xis)

(m)

Fig. 7. Difference between the giver and the receiver position in the zaxis during the passing phase. As we can see the giver and receiver moverelative to each other as they transfer the object.

in the z-axis as a function of the load share, fitted with aGaussian Process (GP) regression for visualization purposes.Interestingly, the load transfer does not happen from a con-stant position: at the beginning of the passing phase (αg ≈ 1)the giver’s position is higher, while at the end (αg ≈ 0) thereceiver’s position is higher. These results show that thereis motion during the passing phase of handovers: the giverlowers their hand and the receiver raises their hand, indicatingthat humans rely in part on arm motion in order to performhandovers.

V. A FLUID HUMAN-INSPIRED HANDOVER CONTROLLER

Exploiting the insights and data from the study from theprevious section, in this section we detail the motion designof our proposed handover controller, the estimation of thehandover location, and the low-level control necessary toperform fast bi-directional handovers.

A. Desired trajectories

In our analysis we observed that during the approachphase both the giver and the receiver exhibit a goal-orientedbehavior towards the handover pose. In addition, the giver’sposition leads the motion while the receiver’s motion (andother variables such as the grip force) exhibit a subordinatebehavior. Concerning the passing phase, our analysis indi-cates that it is driven by the perceived load share of eachparticipant, which highlights the importance of arm motionduring this phase. Coupled Dynamical System (CDS) [13]are well-suited to capture these constraints a priori, where aslave DS evolves depending on the state of the master DS.

Concerning the giver’s arm position, we assume the fol-lowing structure

pg = fg(pg − gg(αg)) . (8)

where fg is a continuous and continuously differentiablefunction with a single attractor given by gg(αg). In addition,we consider a priori that gg(αg = 1) = pg,ho. This way,during the approach phase, the giver follows the attractorgiven by the handover location, i.e. pg = fg(pg − pg,ho), asan autonomous system. After contact with the other party, αgshifts the attractor as observed in Section IV-B3. This for-mulation inherently assumes that the load share αg is themaster DS driving the advancement of the task, as highlightedin Section IV-B2 and g(αg) represents the coupling functionbetween the motion of the giver and the load share.

From the insights of Section IV-B1, the receiver’s armmotion is given by a slave DS that depends on the giver’smotion and the load share

pr = fr(pr − gr(z)) , (9)

where

z =

{‖pg − pg,ho‖, if αg = 1 (approach)αg − 1, if 0 ≤ αg ≤ 1 (passing)

Again, fr is a continuous and continuously differentiablefunction with a single attractor given by gr(z). Note that thecoupling function gr(z) is expressed in term of the auxiliaryvariable z, which encompasses both the advancement of thegiver’s arm during the approach phase and the transfer ofthe load during the passing phase. The end-effector angularmotion of both parties is coupled to z in a similar manner.

To synthesize desired hand motions, instead of modelinghand configurations xh,g,xh,r directly, we modulate the gripforce during the handover with a slave DS that depends onthe auxiliary variable z

ui = fh,i(uh,i − gh,i(z)). (10)

We characterize the handover dynamics in Equations (8)to (10) using data gathered in our human-human handoverstudy. We learn the giver’s arm position dynamics (Equa-tion (8)) by estimating the joint distributions P (pg,pg)and P (αg,pg) with two Gaussian Mixture Models inferred

Fig. 8. Depiction of the receiver’s estimation of the handover pose for twodifferent trajectories of the giver (in blue). We model the giver’s motion asa linear dynamical system, and predict the handover pose as the one closestto the receiver along the predicted trajectory.

from our study data. Functions gg and fg are given by theexpected values of the respective conditional probabilities:

gg(αg) = E[P (pg|αg)]fg(pg − gg(αg)) = E[P (pg|(pg − gg(αg)))] .

Estimating the remaining dynamics (Equations (9) and (10))is done in a similar fashion. For a more detailed explanationwe refer the reader to [13].

B. Low-level control

To reproduce the desired trajectories of the arm on atorque-controlled manipulator while at the same time en-suring a compliant behavior during the passing phase, weemploy an impedance-based control scheme

τ = JT(Kd(xd,i − xi) + αig) + τ ext ,

where τ is the torque input, τ ext are the external torques, Jis the Jacobian, Kd is a velocity-tracking gain and xd,i withi ∈ {g, r} is the desired twist of the i-th party given by thelearned translational velocities (8)(9) and angular velocities.The term αig compensates for the proportion of the object’sweight supported by the robot.

To simplify the control of a humanoid hand, we leveragerecent work in grasp synergies [14] to project the high-dimensional control problem into a lower-dimensional sub-space. We modulate the desired grip force ui by jointlymodeling the hand-object around the contact point as aspring in synergy space [4]. A simple mapping from desiredgrip force to a synergy weighting (for instance, scalingthe first principal component) results in a grasp that scalesaccording to the desired grip force. While in principle thismapping is object-dependent, in our studies we only considerobjects of similar size. Furthermore, for simplicity we ensurethis mapping provides a conservative estimate of the forcenecessary to successfully grasp any object we hand over.Object-specific mappings (as well as mappings that vary overdifferent object grasp points) remain as future work.

C. Handover pose estimation

Our handover controller formulation requires an estimateof the handover pose (the pose where both parties meet) inadvance. We model the motion of the human as a dynamicalsystem and use it to predict the handover pose continuously.

Fig. 9. We evaluated our controller on a KUKA LWR robot equipped withan Allegro hand and the ATI Gamma force/torque sensor, with an Optitrackmarker on the human’s wrist. Our approach works with any graspable objectwith a known mass.

Using the pose data of the human for a pre-defined timewindow, we estimate a linear dynamical system xi = Axiusing least squares approximation. Using the model, thehuman hand motion is predicted and the position closest tothe robot is estimated as the current handover location. Theestimate evolves over time, and converges to the human pose,assuming that it is in the workspace of the robot. A schematicof finding the handover estimate is shown in Figure 8.

VI. HUMAN-ROBOT HANDOVER EXPERIMENTS

To validate our proposed controller we implemented it ona robotic platform and compared it to a baseline controllerbased on thresholds.

1) Experimental setup: We implemented our human-inspired controller for fluid handovers on a 7-DoF KUKALWR 4+ robot, equipped with the SimLabs Allegro Hand (a16-DoF humanoid hand). The robot is also equipped with aATI Gamma force/torque sensor mounted at the wrist, whichmeasures the interaction force on the object as it is passedfrom the giver to the receiver. We use an OptiTrack systemto measure the 6-DoF pose of the participant’s wrist.

2) Experimental procedure: We evaluate the controller inhandovers where the robot is the giver. The robot begins withthe object in its hand. For the evaluation we used the plasticwater bottle shown in Figure 9 weighting mo = 0.556 [kg].The specific grasp is not important; in our experiments wehave the robot close its hand on the object, resulting in apower grasp of the object. We perform a calibration step withthe robot is a static configuration to compute the baselineforces/torques from the object; this accounts for the massof the hand and object. We then begin the handover with asignal to the human subject to move towards the robot toreceive the object. We consider the handover complete whenthe robot no longer has the object in its hand.

3) Experimental conditions: We evaluate two differenthandover controllers on the robot:• Our fluid human-inspired controller (described in Sec-

tion V) which continuously estimates the handover poseand moves the robot end-effector towards it. The con-troller simultaneously monitors the load share estimate,

using it to control the robot hand’s grip force. Notethat thanks to our decoupled formulation, the robot canstill be moving when it begins to release the grasp, forexample if an unknown third party should take the objectas the robot is moving.

• A threshold-based baseline controller which switchesbetween discontinuous phases. When the robot is thegiver, it first applies a constant grip force (the same asin our controller) to grasp the object. Then the robotapproaches the handover pose and stops at a distanceof 0.1 [m] from it. The robot waits until the sensed loadshare is lower than a hand-tuned threshold of αg = 0.2and then completely opening the hand in an open-loopfashion. The switching-based controller for giving theobject follows the same idea. Note that since the loadshare is computed directly from the wrist force/torquesensor, in the absence of forces due to acceleration, theload share threshold approach is equivalent to simplydetecting changes in the raw force readings along thevertical axis.

For each of these two controllers, we recorded 4 handoversperformed with a single healthy subject.

4) Measures: To evaluate the performance of the han-dovers we consider two measures:• The internal wrench norm between the robot and the

human uint computed as in (2). Internal wrenchesrepresent counteracting wrench components which areunnecessary to accomplish the task.

• The duration of the passing phase, defined as the timeduring which both parties are supporting the object.We compute this duration as in [1, 2, 3]: the pass-ing phase is the time during the handover where theload share is shared between the giver and receiver(i.e. εα < αg < 1− εα for a threshold εα).

Since internal wrenches represent forces on the object thatserve no purpose towards the completion of the handover,they can be seen as nonessential, and thus a lower inter-nal wrench norm is desirable to enable natural handovers.Similarly, a faster handover is desirable since both partiesspend less time in contact with the object negotiating theload transfer.

A. Results

Our handover experiments indicate that our controllersignificantly reduces internal forces between the robot andthe human, compared to the threshold-based controller. Oneinstance of the handover dynamics for two handovers isshown in Figure 10. As we can see, waiting for the loadshare to reach a threshold before opening results in highinternal forces on the object, whereas actively controlling thegrip force and the arm using the sensed load share results ina more natural handover that remains in motion during thepassing phase.

The average internal force norm and duration of thepassing phase across trials are shown in Figure 11. Ourproposed controller reduces internal forces w.r.t the threshold-based controller both in terms of the average of means and

0

0.5

1αg

Load share

0

4

8

‖uin

t‖ Internal wrench

0.0 0.25 0.5 0.75 10

0.1

0.2

Time (s)

‖pg‖ Robot velocity

(a) Our fluid handover controller.

0

0.5

1

αg

Load share

0

4

8

‖uin

t‖ Internal wrench

0.0 0.25 0.5 0.75 10

0.1

0.2

Time (s)

‖pg‖ Robot velocity

(b) A threshold-based baseline controller.

Fig. 10. Dynamics of robot-to-human handovers with our strategy (10a) and a threshold-based strategy (10b). As we can see, in both cases the giver’sload share (αg) decreases during the handover, whereas the magnitude of the internal wrench force (‖uint‖) is much lower for our controller since weactively control the grip force (leading to a smoother change in load share). Note that since our controller formulation does not depend on a discrete phaseswitch, the robot is always in motion (‖pg‖ > 0), even during the object transfer.

Mean Max0

2

4

6

8

10

Inte

rnal

Forc

e(N

)

Passing Phase

0.1

0.2

0.3Ti

me

(s)

Our ApproachThreshold-based

Fig. 11. Comparison of internal forces and duration of the passing phasebetween our fluid controller and a threshold-based controller. As we can see,our controller yields faster handovers with lower internal force, resulting ina more natural handover.

the average of maximum values. This results in handoverswhich require less unnecessary work. The duration of thepassing phase is also significantly reduced. Continuouslycontrolling the arm and the hand during the handover yieldsa more responsive interaction and a faster passing of theobject as shown in the video available at https://youtu.be/Ac4kgipC7A0.

VII. CONCLUSIONS

In this work we presented a novel human-inspired bidi-rectional human-robot handover controller. Its design is sup-ported by insights from existing studies as well as novelinsights observed in our human-human experiments: ourresults indicate the existence of motion during the passingphase as well as a coupling between the motions of thegiver and the receiver. Our controller is based on a phase-less handover dynamics model that produces smooth and fasthandovers. Experiments show that our controller is smooth,fast, and reduces internal forces on the object compared totraditional switching-based approaches.

ACKNOWLEDGMENTS

We gratefully acknowledge the funding provided by the EuropeanCommission (Horizon 2020 Framework Programme, H2020-ICT-643950)for the SecondHands project. Thanks to Klas Kronander and GuillaumedeChambrier for their help with the robot experiments.

REFERENCES[1] A. H. Mason and C. L. MacKenzie, “Grip forces when passing an

object to a partner,” Experimental Brain Research, 2005.[2] W. P. Chan, C. A. Parker, H. M. Van der Loos, and E. A. Croft,

“Grip forces and load forces in handovers: Implications for DesigningHuman-Robot Handover Controllers,” in Human-Robot Interaction,2012.

[3] S. Endo, G. Pegman, M. Burgin, T. Toumi, and A. M. Wing, “Hapticsin between-person object transfer,” in Haptics: Perception, Devices,Mobility, and Communication, 2012.

[4] W. P. Chan, C. A. Parker, H. M. Van der Loos, and E. A. Croft, “Ahuman-inspired object handover controller,” The International Journalof Robotics Research, 2013.

[5] M. Cakmak, S. S. Srinivasa, M. K. Lee, J. Forlizzi, and S. Kiesler,“Human Preferences for Robot-Human Hand-over Configurations,” inInternational Conference on Intelligent Robots and Systems, 2011.

[6] J. Waldhart, M. Gharbi, and R. Alami, “Planning handovers involvinghumans and robots in constrained environment,” in InternationalConference on Intelligent Robots and Systems, 2015.

[7] J. Mainprice, M. Gharbi, T. Simeon, and R. Alami, “Sharing effort inplanning human-robot handover tasks,” in Proceedings of RO-MAN,2012.

[8] V. Micelli, K. Strabala, and S. S. Srinivasa, “Perception and controlchallenges for effective humanrobot handoffs,” in RSS 2011 RGB-DWorkshop, 2011.

[9] M. Prada, A. Remazeilles, A. Koene, and S. Endo, “Dynamic Move-ment Primitives for Human-Robot interaction: Comparison with humanbehavioral observation,” in International Conference on IntelligentRobots and Systems, 2013.

[10] S. Erhart and S. Hirche, “Internal force analysis and load distributionfor cooperative multi-robot manipulation,” IEEE Trans. Robot., vol. 31,no. 5, pp. 1238–1243, 2015.

[11] C. Granger, “Some recent development in a concept of causality,”Journal of Econometrics, 1988.

[12] L. Barnett and A.Seth, “The MVGC multivariate granger causalitytoolbox: A new approach to granger-causal inference,” Journal ofNeuroscience Methods, vol. 223, pp. 50–68, 2014.

[13] A. Shukla and A. Billard, “Coupled dynamical system based arm-handgrasping model for learning fast adaptation strategies,” Robotics andAutonomous Systems, vol. 60, no. 3, pp. 424 – 440, 2012.

[14] M. Santello, M. Flanders, and J. F. Soechting, “Postural hand synergiesfor tool use,” The Journal of Neuroscience, vol. 18, no. 23, pp. 10 105–10 115, 1998.